Options
Advanced feature set and classification model for heart sound signal analysis
Date Issued | Volume | Issue | Start Page | End Page |
---|---|---|---|---|
2024-04-05 | 60 | Suppl. 1 | 151 | 151 |
Background. Congenital heart defects are the important morbidity and mortality factors for several severe clinical conditions. This birth defect is detected and evaluated by experts using heart auscultation. However, despite the reported quite good agreement between the experts about the general diagnosis, there are no clearly described features or quantitative parameters which can be used for classification. Cardiologists are ‘looking for’ and subjectively grading the audibility of whooshing or swishing sounds heard during the heartbeat. The sought sounds indicate that blood is flowing abnormally across heart valves due to structural heart problems. Machine learning models trained with optimal feature sets should be capable of classifying phonocardiogram signals into ‘reflecting heart defect’ and ‘normal’ ones. By implementing such a classification model into mobile devices or cloud-based systems, we can realize rapid and robust screening of disorders of the mechanical function of the heart in the paediatric population could be realized. Aim. To elaborate optimal phonocardiogram feature set and classification model to detect disorders of the mechanical function of the heart. Methods. We used annotated phonocardiogram recordings of 1568 patients collected from paediatric population screening campaigns conducted in Northeast Brazil in July–August 2014 and June–July 2015, provided by Physionet’s open clinical database. Convolutional neural network elaboration strategy, usually used for image classification or recognition, was applied for the classification of 2-dimensional arrays of time-frequency estimates of signal recording excerpts, constructed by means of continuous wavelet transform and special data pre-processing. Results. 5-fold cross-validation of elaborated algorithm performance showed an accuracy of 0.777 ± 0.03. The accuracy of the algorithm tested with ‘never seen’ data was 0.671. The algorithm was evaluated in the Physionet 2022 challenge and ranked in 23rd place out of 44 participants by classification performance. At the same time, the algorithm required comparatively low computational resources. Conclusion. Due to comparatively low required computational resources and good classification accuracy, the elaborated algorithm is suitable for deployment into mobile diagnostic devices or cloud-based systems for paediatric populational screening.